Multi-radar point cloud weighted fusion system and method based on entropy weight method

Abstract

The application discloses a multi-radar point cloud weighted fusion system and method based on an entropy weight method, belongs to the technical field of intelligent stereoscopic warehouse perception and point cloud fusion, and aims to solve the problems of low fusion precision, poor real-time performance and unreasonable weight distribution of existing multi-radar point cloud fusion. The method comprises the following steps: synchronously collecting and pre-processing multi-radar point clouds to generate a set; separating dynamic and static points to obtain static point clouds; dividing fusion voxels and detecting changed voxels; extracting features and standardizing; combining prior weights and entropy weight methods to calculate fusion weights; and weightedly fusing to generate a fusion point cloud map. The application can significantly improve the precision and real-time performance of intelligent stereoscopic warehouse point cloud fusion, and provides a reliable three-dimensional perception basis for upper-layer operations of a warehouse.

Description

Technical Field

[0001] This invention relates to the field of intelligent automated warehouse sensing and point cloud fusion technology, and more specifically to a multi-radar point cloud weighted fusion system and method based on entropy weighting. Background Technology

[0002] As the core carrier of modern intelligent warehousing systems, intelligent automated warehouses are widely used in operational scenarios such as intelligent storage and retrieval of goods, automated scheduling, and full-area environmental monitoring. To achieve refined three-dimensional perception of the entire warehouse space, the industry generally deploys multiple LiDARs to collaboratively collect point cloud data. Multi-radar point cloud fusion technology has become a key technology direction to ensure the perception effect of warehouse space. Existing multi-radar point cloud fusion modes in warehousing scenarios mostly adopt fixed weights or purely manual experience-based weight allocation methods, which cannot adapt to the differentiated perception needs of different operational areas in the warehouse. At the same time, they are difficult to effectively avoid the negative impact of dynamic interference on static background fusion, generally suffering from low point cloud fusion accuracy, insufficient real-time data processing, and poor stability of fusion results. This makes it difficult to meet the actual usage requirements of high-precision, high-reliability three-dimensional perception data for upper-level applications such as intelligent automated warehouse location status recognition, automated equipment navigation, and safety risk monitoring. Therefore, to overcome these limitations, this invention proposes a multi-radar point cloud weighted fusion system and method based on the entropy weight method. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide a weighted fusion system and method for multi-radar point clouds based on the entropy weight method. This solves the problems of low fusion accuracy, poor real-time performance, unscientific weight allocation, and unreliable fusion results due to dynamic interference in existing intelligent automated warehouse multi-radar point cloud fusion.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A weighted fusion method for multi-radar point clouds based on entropy weighting includes:

[0006] In response to the acquisition trigger signal of the intelligent automated warehouse, the point cloud data of all LiDARs in the current frame are acquired synchronously, point cloud data preprocessing is performed, and a point cloud data set of the current frame is generated.

[0007] Perform independent temporal dynamic and static separation on the current frame point cloud data set using a single LiDAR, distinguish between dynamic and static points, and generate a static point cloud data set;

[0008] Based on the functional partitioning of the fusion voxel mesh, the current frame static point cloud data set is subjected to single-liDAR independent adjacent frame change detection to distinguish changing voxels from static voxels and generate a set of voxels to be processed.

[0009] For each variable voxel in the set of voxels to be processed, the multi-dimensional quality features of each lidar within the variable voxel are extracted and differentiated and standardized to obtain the quality feature scores of each lidar within the variable voxel.

[0010] Based on the functional partition of the variable voxel, a combined weight is constructed by combining prior weight and entropy weight method. The comprehensive score of each lidar is calculated based on the quality feature score. After normalization, the comprehensive fusion weight of each lidar in the corresponding variable voxel is obtained.

[0011] Based on the comprehensive fusion weight, an incremental voxel-level weighted fusion operation is performed on the changing voxels, and the fused point cloud map of the current frame is generated by combining the fusion result of the static voxels from the previous frame.

[0012] Specifically, the steps for generating a static point cloud dataset include:

[0013] Initialize the background difference voxel and the background difference counter; the background difference counter is used to count the number of frames in which the point cloud appears consecutively in the corresponding background difference voxel of the corresponding lidar;

[0014] For each lidar, iterate through all discrete points in the current frame point cloud data set, determine the background difference voxel to which the discrete point belongs based on its three-dimensional coordinates, and update the value of the corresponding background difference counter.

[0015] Based on the value of the background difference counter and the preset counting threshold, the discrete points within each background difference voxel are determined to be static or dynamic points.

[0016] All discrete points identified as static points by LiDAR are integrated to generate a static point data set for the current frame.

[0017] Specifically, the steps for generating the set of voxels to be processed include:

[0018] Based on the functional zoning of the intelligent automated warehouse, differentiated voxel grid division is carried out;

[0019] Obtain the fusion result data of the previous frame's static point cloud data set and the previous frame's fusion result data;

[0020] For each LiDAR, its current frame static point cloud data set is compared with the previous frame static point cloud data set according to the fused voxel grid, and the change of each fused voxel is calculated.

[0021] Based on the preset change threshold and the calculated change amount of the fused voxel, mark the changed voxel detected by each lidar;

[0022] A multi-radar voting mechanism is used to merge all the variable voxels detected by lidar, generating a variable voxel set and a static voxel set, and the variable voxel set is used as the voxel set to be processed.

[0023] Specifically, the changes in each fused voxel include changes in coordinates and changes in the number of points;

[0024] The step of calculating the change in each fusion voxel includes:

[0025] For each LiDAR, all discrete points in its current frame static point cloud data set are assigned to the corresponding fusion voxels according to their three-dimensional coordinates;

[0026] For each non-empty fused voxel, calculate the average coordinates and number of points of the static point cloud within the fused voxel in the current frame, and the average coordinates and number of points of the static point cloud within the fused voxel in the previous frame.

[0027] The coordinate change is calculated based on the distance between the average coordinates of the static point cloud within the fused voxel in the current frame and the average coordinates of the static point cloud within the fused voxel in the previous frame.

[0028] The change in the number of points is calculated based on the absolute difference between the number of static points in the fused voxel in the current frame and the number of static points in the fused voxel in the previous frame.

[0029] Specifically, the steps of extracting multi-dimensional quality features of each lidar within a variable voxel and performing differential standardization to obtain the quality feature scores of each lidar within that variable voxel include:

[0030] All discrete points in the current frame static point cloud data set are assigned to corresponding variable voxels according to their three-dimensional coordinates, and a point cloud mapping relationship between each lidar and each variable voxel is established.

[0031] Based on the point cloud mapping relationship, multi-dimensional quality features of each lidar in each changing voxel are extracted. The multi-dimensional quality features include basic geometric features and echo features.

[0032] All extracted quality features are classified into positive features, negative features, and moderate features according to their physical meaning. Differential standardization processing is performed, and they are uniformly mapped to the [0,1] interval to obtain the quality feature scores of each LiDAR within the variable voxel.

[0033] Specifically, the basic geometric features are used to reflect the geometric perception accuracy of the lidar within the variable voxel, including point cloud density, plane fitting error, average ranging distance, and average incident angle.

[0034] The point cloud density is the ratio of the number of discrete points of the lidar within the variable voxel after preprocessing to the volume of the variable voxel.

[0035] The plane fitting error is the root mean square error of fitting the discrete points of the lidar under the current changing voxel within the point cloud mapping relationship to the point cloud plane.

[0036] The average ranging distance is the arithmetic mean of the distances from all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship to the origin of the lidar's own coordinate system.

[0037] The average incident angle is the arithmetic mean of the angles between the laser incident direction and the normal vector of the point at all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship.

[0038] Specifically, echo characteristics are used to reflect the quality of the echo signal of the lidar within the variable voxel, including average echo intensity, number of echoes and echo interval;

[0039] The average echo intensity is the arithmetic mean of the echo intensity of all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship.

[0040] The number of echoes is the maximum value of the echo numbers of all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship;

[0041] The echo interval is the arithmetic mean of the time differences between adjacent echo points of the lidar under the current changing voxel within the point cloud mapping relationship.

[0042] Specifically, the calculation steps for the comprehensive fusion weight of each lidar within the corresponding variable voxel include:

[0043] Obtain the prior weights of the quality features corresponding to the functional partitions;

[0044] Based on the quality feature scores of all lidars within the current changing voxel, the data dispersion and information contribution of each quality feature are quantified, and the objective entropy weight of each quality feature is calculated.

[0045] The prior weights and objective entropy weights are weighted and fused according to a preset weight ratio to obtain the combined weights of each quality feature.

[0046] For each lidar, its scores for each quality feature within the current changing voxel are multiplied sequentially by the combined weights of the corresponding quality features and then summed to obtain the comprehensive score of the lidar.

[0047] The overall scores of each lidar are globally normalized to obtain the overall fusion weight.

[0048] Specifically, the steps for generating the fused point cloud map of the current frame include:

[0049] For each variable voxel, extract the static point cloud data of each participating fusion lidar within that variable voxel from the point cloud mapping relationship, including the three-dimensional coordinates and echo intensity information of each discrete point;

[0050] Based on the comprehensive fusion weight, the three-dimensional coordinates and echo intensity are weighted and fused separately to obtain the fused coordinates and fused intensity of the changed voxel;

[0051] The fusion coordinates and fusion intensity of all the variable voxels are stitched together with the fusion coordinates and fusion intensity of the static voxels from the previous frame to obtain the fused point cloud map of the current frame.

[0052] Specifically, the steps for generating the current frame point cloud data set include:

[0053] Responding to the data collection and triggering signals of the intelligent automated warehouse;

[0054] Based on a unified time reference, the current frame point cloud data of all lidars are collected synchronously.

[0055] Each lidar-collected point cloud data is preprocessed independently. The preprocessing operations include removing outlier data, filtering out noise data, and performing motion compensation on the point cloud data.

[0056] All preprocessed point cloud data from LiDAR are uniformly converted to the global coordinate system of the intelligent automated warehouse and integrated to generate the current frame point cloud data set.

[0057] A multi-radar point cloud weighted fusion system based on entropy weighting includes:

[0058] Data acquisition module: In response to the acquisition trigger signal of the intelligent automated warehouse, it synchronously acquires the point cloud data of all LiDARs in the current frame, performs point cloud data preprocessing, and generates the point cloud data set of the current frame;

[0059] The static / dynamic separation module is used to perform independent temporal dynamic / static separation of the current frame point cloud data set by a single LiDAR, distinguishing between dynamic and static points, and generating a static point cloud data set.

[0060] Change detection module: Based on the functional partitioning of fused voxel mesh, it performs single-LiDAR independent adjacent frame change detection on the current frame static point cloud data set, distinguishes between changing voxels and static voxels, and generates a set of voxels to be processed;

[0061] Feature extraction module: For each variable voxel in the set of voxels to be processed, extract the multi-dimensional quality features of each LiDAR within the variable voxel, and perform differential standardization processing to obtain the quality feature scores of each LiDAR within the variable voxel.

[0062] Weight calculation module: Based on the functional partition to which the variable voxel belongs, a combined weight is constructed by combining prior weight and entropy weight method. The comprehensive score of each lidar is calculated based on the quality feature score. After normalization, the comprehensive fusion weight of each lidar in the corresponding variable voxel is obtained.

[0063] Point cloud map generation module: Based on comprehensive fusion weights, it performs incremental voxel-level weighted fusion operations on changing voxels, and combines the fusion results of the previous frame with the static voxels to generate the fused point cloud map of the current frame.

[0064] The beneficial effects of this invention are:

[0065] This invention eliminates dynamic point interference during the operation of an intelligent automated warehouse by independently separating and statically fusion data from a single lidar. It accurately identifies the changing voxels to be processed by functional zoning and differentiated voxel division. It comprehensively quantifies the perception advantages and disadvantages of each lidar based on multi-dimensional geometric and echo quality characteristics. Then, it integrates functional zoning prior weights and entropy weighting to construct a combined weight that considers both subjective and objective factors. Combined with incremental voxel-level weighted fusion that reuses historical results of static voxels, this invention effectively adapts to the differentiated perception needs of different areas of the warehouse while significantly reducing redundant computation, improving the real-time performance of point cloud processing, and accurately highlighting the fusion contribution of high-reliability lidar data. This significantly improves the accuracy and stability of static background point cloud fusion across the entire intelligent automated warehouse. Attached Figure Description

[0066] Figure 1 This is a flowchart of the multi-radar point cloud weighted fusion method based on entropy weighting method of the present invention;

[0067] Figure 2 This is a flowchart illustrating the process of obtaining the set of voxels to be processed according to the present invention;

[0068] Figure 3 This is a flowchart for obtaining quality feature scores in this invention;

[0069] Figure 4 This is a flowchart for obtaining the comprehensive fusion weight in this invention. Detailed Implementation

[0070] Example 1

[0071] Please see Figure 1 This embodiment introduces a multi-radar point cloud weighted fusion method based on entropy weighting, which is applied to an intelligent automated warehouse. The intelligent automated warehouse includes at least two lidar units for collecting point cloud data. The point cloud data refers to a set of three-dimensional discrete points generated by the lidar by emitting laser pulses and receiving echoes. Each discrete point contains three-dimensional coordinates, echo intensity, timestamp, and echo number information. The method includes:

[0072] Step S1: In response to the acquisition trigger signal of the intelligent automated warehouse, synchronously acquire the point cloud data of all LiDARs in the current frame, perform point cloud data preprocessing, and generate the point cloud data set of the current frame;

[0073] Furthermore, step S1 includes:

[0074] Step S11: Respond to the acquisition trigger signal of the intelligent automated warehouse, the acquisition trigger signal being used to trigger all LiDARs in the intelligent automated warehouse to synchronously perform point cloud data acquisition operations;

[0075] The trigger signal can be automatically generated according to a preset timing period or manually triggered by the intelligent automated warehouse control system. The preset timing period is set according to the dynamic operation level and global perception accuracy requirements of the intelligent automated warehouse. When the warehouse has frequent inbound and outbound operations, the timing period can be shortened to increase the perception frequency; when the warehouse is in a static storage state, the timing period can be extended to reduce system power consumption. For example, the preset timing period is preferably 1 Hz. The coverage of the timing trigger signal is the entire spatial area of ​​the intelligent automated warehouse, which is used to realize the periodic monitoring of the global status of the warehouse and maintain the basic perception capability of the system when there is no task trigger signal.

[0076] Step S12: Based on a unified time reference, synchronously collect the current frame point cloud data of all LiDARs; establish a unified time reference for multiple LiDARs, synchronously collect point cloud data, and respond to the collection trigger signal generated in step S11, all LiDARs start scanning at the same time to collect the point cloud data of the current frame; at least two LiDARs are deployed in the intelligent automated warehouse, for example, four 32-line mechanical rotating LiDARs, which are installed on the ceiling in the four corners of the intelligent automated warehouse, with an installation height of 6 meters, a field of view of 360 degrees × 90 degrees, and a scanning frequency of 10 Hz. A single LiDAR can collect approximately 32,000 raw point cloud data per frame, and the four LiDARs collect approximately 128,000 point cloud data per frame.

[0077] Step S13: Perform preprocessing operations independently on the point cloud data collected by each lidar. The preprocessing operations include removing outlier data, filtering out noise data, and performing motion compensation on the point cloud data.

[0078] Furthermore, step S13 includes:

[0079] Step S131: Based on the three-dimensional coordinates and echo intensity of each discrete point in the point cloud data, outlier data is removed. These outlier data include distance anomalies and intensity anomalies. Distance anomalies with distances less than a preset minimum ranging threshold or greater than a preset maximum effective ranging threshold are removed. The preset minimum ranging threshold and preset maximum effective ranging threshold are determined based on the technical parameters of the LiDAR used. When the distance is less than the minimum ranging threshold, the ranging accuracy of the LiDAR will significantly decrease; when the distance is greater than the maximum effective ranging threshold, the echo signal strength is insufficient, easily generating erroneous data. For example, the preset minimum ranging threshold and the preset maximum effective ranging threshold are 0.5 meters and 50 meters, respectively; remove intensity anomalies where the echo intensity is less than the preset minimum intensity threshold or greater than the preset maximum intensity threshold. The preset minimum intensity threshold and the preset maximum intensity threshold are determined based on the echo signal characteristics of the lidar. Points with an echo intensity less than 10 are usually weak echoes generated by low-reflectivity materials or environmental noise, while points with an echo intensity greater than 250 are usually saturated signals generated by specular reflection. For example, the preset minimum intensity threshold and the preset maximum intensity threshold are 10 and 250, respectively.

[0080] Step S132: Perform statistical filtering denoising operation, using statistical filtering to remove outlier noise points from the original point cloud data; calculate the average distance from each discrete point to its preset number of nearest neighbor discrete points, the preset number being set according to the average density of the point cloud, for example, the preset number is 20; calculate the mean and standard deviation of the average distances of all discrete points, and remove outlier noise points whose average distance is greater than the mean plus a preset multiple of the standard deviation, the preset multiple being set according to the noise level of the point cloud. According to the statistical characteristics of normal distribution, approximately 95.4% of the sample data will fall within the range of mean ± 2 times the standard deviation, and points exceeding this range can be identified as outlier noise points; this multiple can effectively remove the vast majority of noise while retaining the true point cloud data; for example, the preset multiple is preferably 2 times;

[0081] Step S133: If the lidar is installed on a mobile device, such as a four-way shuttle, an automated guided vehicle, or a high-level stacker crane, motion compensation is required to correct motion distortion. Angular velocity and acceleration data of the mobile device where the lidar is located are collected. The pose change of the lidar within a single scan cycle is calculated by integration. The three-dimensional coordinates of each discrete point are corrected based on the pose change to eliminate point cloud distortion caused by the lidar's motion. The single scan cycle refers to the time required for the lidar to complete a full 360-degree rotation scan and generate one frame of point cloud data, and its value is equal to the reciprocal of the lidar's scanning frequency. The pose change refers to the three-dimensional translation and rotation of the lidar relative to its initial position within a single scan cycle. This can be obtained by integrating the angular velocity and acceleration data collected by the inertial measurement unit rigidly connected to the lidar, or by obtaining the odometer data from the mobile device itself. These pose acquisition methods are existing technologies and will not be elaborated upon here. For fixed-installation lidars, since there is no motion distortion, this step is skipped.

[0082] Specifically, step S13 aims to purify the point cloud data, improve its quality, and ensure high accuracy and reliability of the point cloud data used in subsequent processing. In intelligent automated warehouses, during the acquisition process, the LiDAR's ranging accuracy is significantly reduced at close range due to its limited range, while weak echo signals at long distances easily generate erroneous data. Furthermore, the different reflective characteristics of materials such as insulation layers and metal components can lead to abnormal echo intensity. If these abnormal points are not removed, they will interfere with subsequent voxel segmentation and feature extraction. The point cloud data acquired by the LiDAR may also contain isolated outlier noise points due to environmental interference and equipment errors. These noise points will generate false signals in subsequent gradient calculations and change detection, leading to misjudgments. If the LiDAR is installed on mobile devices such as four-way shuttles or automated guided vehicles, the movement of the LiDAR during scanning will cause distortion in the point cloud, causing the three-dimensional coordinates of discrete points to deviate from their actual positions, affecting the accuracy of spatial positioning. Step S13 removes distance and intensity outliers through S131 to ensure the validity of the point cloud data; through statistical filtering denoising in S132, outlier noise points are effectively removed while retaining the true point cloud data, avoiding noise interference; through motion compensation in S133, distortion correction is performed for LiDAR on mobile devices to eliminate coordinate deviations caused by motion. This step is skipped for fixed LiDAR to balance processing accuracy and efficiency. If step S13 is missing, outliers and noise points will be mixed into subsequent processing flows, resulting in inaccurate voxel segmentation, excessive feature extraction deviation, and motion distortion that will distort the spatial position of the point cloud, ultimately affecting the accuracy of the fused point cloud map and failing to meet the perception requirements of intelligent automated warehouses.

[0083] Step S14: Convert all preprocessed point cloud data from the LiDAR to the global coordinate system of the intelligent automated warehouse;

[0084] The extrinsic parameters of each LiDAR relative to the global coordinate system are obtained in advance. The extrinsic parameters include the rotation matrix and the translation vector, which are obtained in advance by the hand-eye calibration method. Based on the extrinsic parameters, the preprocessed point cloud data of each LiDAR is transformed from its own local coordinate system to a unified global coordinate system, so that the point cloud data of all LiDARs are in the same spatial reference, providing a unified spatial basis for subsequent voxel division, dynamic and static separation and fusion operations.

[0085] Step S15: Integrate the preprocessed point cloud data of all lidars converted to the global coordinate system to generate the current frame point cloud data set; the current frame point cloud data set contains all discrete point data collected and preprocessed by all lidars in the current frame, and each discrete point retains its original three-dimensional coordinates, echo intensity, timestamp, echo number information and the identification information of the lidar to which it belongs; the current frame point cloud data set serves as the unified input for all subsequent processing steps, ensuring that subsequent steps can access the complete observation data of all lidars.

[0086] Specifically, step S1 is the foundational data acquisition and preprocessing stage of the entire multi-radar point cloud weighted fusion method. Its core design addresses key issues in intelligent automated warehouses, such as asynchronous acquisition by multiple lidars, poor quality of raw point cloud data, and inconsistent coordinate systems. This provides high-quality, standardized point cloud data input for subsequent dynamic / static separation, change detection, and fusion operations. In intelligent automated warehouse scenarios, multiple lidars are deployed in different locations, and their activation timing may vary. Without a synchronized acquisition mechanism, the point cloud data from different lidars will be out of sync, failing to accurately reflect the spatial state of the warehouse at any given moment. Furthermore, the raw point cloud data acquired by lidars inevitably contains outliers and noise points; if installed on mobile devices, motion distortion will also occur. These interferences severely affect the accuracy of subsequent processing. In addition, each lidar has its own local coordinate system; without unifying it to a global coordinate system, the multi-radar point cloud data cannot be spatially aligned, making fusion operations impossible. Step S1 ensures that all radars acquire data at the same time by controlling the acquisition trigger signal in S11 and synchronizing the unified time reference in S12, thus avoiding timing deviations. The preprocessing operation in S13 removes outliers, filters noise, and corrects motion distortion to purify the point cloud data. The coordinate system transformation in S14 achieves spatial unification of the point clouds from multiple radars. The integration operation in S15 generates a unified set of point cloud data for the current frame, providing a consistent input basis for all subsequent steps.

[0087] Step S2: Perform independent temporal dynamic and static separation of the current frame point cloud data set using a single LiDAR, distinguish between dynamic points and static points, and generate a static point cloud data set;

[0088] Furthermore, step S2 includes:

[0089] Step S21: Initialize the background difference voxel and the background difference counter. The background difference voxel is a three-dimensional spatial cube unit used for temporal background difference. Maintain an independent background difference counter for each background difference voxel and each lidar. The background difference counter is used to count the number of frames in which the point cloud appears consecutively in the corresponding background difference voxel for the corresponding lidar.

[0090] Furthermore, step S21 includes:

[0091] Step S211: Divide the background differential voxel mesh. Divide the global three-dimensional space of the intelligent automated warehouse into multiple non-overlapping background differential voxels. The size of the background differential voxels is set according to the size of the smallest dynamic object in the intelligent automated warehouse and the point cloud density of the LiDAR. When the voxel size is too small, the natural discreteness of the LiDAR point cloud will cause static voxels to occasionally have no points, resulting in excessive counter fluctuations and false judgments. When the voxel size is too large, adjacent dynamic and static objects will be mixed in the same voxel, resulting in missed detections. For example, the size of the background differential voxel is set to 10 cm × 10 cm × 10 cm.

[0092] Step S212: Initialize the background difference counter. Initialize an integer counter with a value range from 0 to the preset maximum value for each background difference voxel and each LiDAR. The preset maximum value is used to prevent the counter of static objects from growing indefinitely and wasting system memory resources. Its value is set according to the statistical time window of the background difference. For example, the preset maximum value is set to 20. When running for the first time, the initial value of all background difference counters is 0.

[0093] Specifically, step S21 is the foundational preparatory step for dynamic-static separation. Its core design is to establish a unified temporal statistical benchmark to address the misjudgment and missed detection issues caused by point cloud discreteness and object size differences, providing reliable spatial units and statistical carriers for subsequent counter updates and point cloud attribute determination. In intelligent automated warehouses, LiDAR point clouds inherently exhibit discreteness. If the background differential voxel size is not set appropriately—for example, if the size is too small, static voxels may occasionally be missing due to point cloud discreteness, leading to excessive counter fluctuations and misjudging static points as dynamic points. If the size is too large, adjacent dynamic and static objects may be mixed into the same voxel, resulting in missed dynamic point detection and an inability to accurately distinguish between the two. Furthermore, if the background differential counter is not initialized and its range limited, the static object counter will grow indefinitely, wasting system memory resources and impacting system efficiency. Step S21 involves dividing the background differential voxel mesh using S211, setting reasonable voxel sizes based on the minimum dynamic object size and point cloud density in the warehouse, ensuring that each voxel can accurately carry point cloud data while effectively distinguishing between dynamic and static objects. Step S212 initializes the background differential counter, assigning an independent counter to each voxel and each radar, setting its value range and initial value to avoid counter overflow and memory waste, and providing a unified benchmark for subsequent time-series statistics. Without step S21, subsequent counter updates would lack a clear spatial carrier and initial benchmark, and the counter values ​​would not accurately reflect the temporal distribution patterns of the point cloud, leading to numerous misjudgments and missed detections in dynamic / static object identification, and preventing the generation of a high-quality static point cloud dataset.

[0094] Step S22: Update the background difference counter. For each LiDAR, iterate through all discrete points in the current frame point cloud data set, determine the background difference voxel to which the discrete point belongs based on its three-dimensional coordinates, and update the value of the corresponding background difference counter.

[0095] Furthermore, step S22 includes:

[0096] Step S221: Determine the background difference voxel to which the discrete point belongs. For each discrete point of each LiDAR, calculate the corresponding background difference voxel index based on its global 3D coordinates. The calculation formula is: , , ,in The global three-dimensional coordinates of the discrete points. Let the side length of the background difference voxel be . This is the floor function;

[0097] Step S222: Update the counter value. For background difference voxels where the LiDAR has discrete points in the current frame, increment the corresponding counter value by 1; for background difference voxels where the LiDAR does not have discrete points in the current frame, decrement the corresponding counter value by 1; when the counter value reaches the preset maximum value, keep the preset maximum value unchanged; when the counter value reaches 0, keep it unchanged at 0; for example, in this embodiment, if a background difference voxel has a discrete point of LiDAR 1 in the current frame, then the counter of the background difference voxel corresponding to LiDAR 1 is incremented by 1, otherwise it is decremented by 1, and the counter value varies between 0 and 20.

[0098] Specifically, step S22 is the core temporal statistics step for dynamic-static separation. Its purpose is to capture the temporal variation patterns of the point cloud within each background differential voxel for each LiDAR unit by updating the background differential counter. This provides a quantitative basis for subsequent dynamic and static point determination, solving the problem of inaccurate attribute determination caused by unclear point cloud temporal distribution. In intelligent automated warehouses, dynamic objects appear in different voxels across different frames, while static objects continuously appear in the same voxel. The increase or decrease of the counter effectively captures this difference. Without a unified voxel index calculation rule, discrete points cannot be accurately assigned to their corresponding voxels, and the counter update will lose its specificity. If the counter update lacks a clear rule and cannot dynamically adjust its value based on the presence or absence of point clouds, the counter value will not accurately reflect the temporal distribution of the point cloud, affecting subsequent determination results. Step S22 calculates the voxel index through S221, accurately determining the background difference voxel to which each discrete point belongs based on the global 3D coordinates of the discrete point and the voxel side length, establishing the correspondence between the point cloud and the voxel. Step S222 updates the counter value, increasing or decreasing the counter based on the presence or absence of the point cloud in the current frame, while limiting the counter's value range to ensure that the counter value accurately and stably reflects the temporal occurrence pattern of the point cloud. Without step S22, the background difference counter cannot be effectively updated, the temporal distribution characteristics of the point cloud cannot be quantified, and the subsequent dynamic and static determination in S23 will lack objective basis, relying solely on subjective experience, leading to inaccurate determination results and a significantly increased probability of missed dynamic points or misjudged static points.

[0099] Step S23: Based on the value of the background difference counter and the preset counting threshold, determine whether the discrete point within each background difference voxel is a static point or a dynamic point. The preset counting threshold is set according to the maximum movement speed of the dynamic object in the intelligent automated warehouse and the scanning frequency of the LiDAR. When the dynamic object moves at its maximum speed, the number of frames that the LiDAR can scan within the time it takes to pass through a single background difference voxel is less than the preset counting threshold, so the counter value will not exceed the preset counting threshold. However, when a static object continuously exists in the voxel, the counter value will stably exceed the preset counting threshold. For example, the maximum movement speed of the dynamic object in the intelligent automated warehouse is 2 m / s, and the scanning frequency of the LiDAR is 10 Hz. The time it takes for an object moving at 2 m / s to pass through a 10 cm background difference voxel is 0.05 seconds, and it will appear in at most 0.5 frames. Considering the discreteness and noise of the point cloud, the dynamic object will appear in the same voxel in at most 2-3 consecutive frames. Therefore, the preset counting threshold is preferably 3.

[0100] When the counter value of a certain background difference voxel corresponding to a certain lidar is greater than a preset counting threshold, all discrete points belonging to that lidar within that background difference voxel are determined to be static points; otherwise, they are determined to be dynamic points.

[0101] Step S24: Separate and transmit point cloud data; merge discrete points identified as dynamic points by all LiDARs to generate a dynamic point dataset for the current frame, track them separately and do not participate in subsequent static background fusion processing; integrate discrete points identified as static points by all LiDARs to generate a static point dataset for the current frame. Each discrete point in the static point dataset retains its original three-dimensional coordinates, echo intensity, timestamp, echo number information, and the identification information of the LiDAR to which it belongs, as input for subsequent change detection steps.

[0102] Specifically, step S2 is a crucial step in separating dynamic and static points. Its core design addresses the interference of dynamic objects on the static background in intelligent automated warehouses, ensuring that subsequent fusion operations are performed only on the static point cloud, thus improving fusion efficiency and accuracy. During the operation of an intelligent automated warehouse, numerous dynamic objects exist. The point cloud data of these dynamic objects changes over time. If they are fused together with the static point cloud, the fused point cloud map will contain dynamic noise, failing to accurately reflect the true state of the warehouse's static background. This also increases the computational load of fusion, affecting the system's real-time performance. Furthermore, multiple lidars have different observation perspectives, and the dynamic / static determination of a single lidar is easily affected by perspective obstruction and sparse point clouds. Independent temporal determination is needed to ensure the reliability of the separation results. Step S2 initializes voxels and counters in S21 to establish a statistical baseline; updates the counters in S22 to capture temporal changes in the point cloud; accurately distinguishes between dynamic and static points based on counter values ​​and preset thresholds in S23, utilizing the temporal distribution differences between dynamic and static objects to ensure the accuracy of the judgment results; and separates dynamic and static points in S24, tracking dynamic points separately and integrating static points into a static point cloud dataset, providing clean input for subsequent change detection and fusion operations. Without step S2, dynamic points would be mixed into the static point cloud data, causing subsequent change detection to misjudge dynamic changes as static background changes. The fusion result would contain a large amount of dynamic noise, failing to provide reliable support for warehouse static background perception, while also increasing the system's computational burden and reducing the real-time performance of the fusion.

[0103] Step S3: Based on the functional partitioning of the fusion voxel mesh, perform single-laser independent adjacent frame change detection on the current frame static point cloud data set, distinguish between changing voxels and static voxels, generate a set of voxels to be processed, and only perform subsequent feature extraction and weight calculation operations on changing voxels, while static voxels directly reuse the fusion result of the previous frame.

[0104] Please see Figure 2 Furthermore, step S3 includes:

[0105] Step S31: Based on the functional zoning of the intelligent three-dimensional warehouse, perform differentiated fusion voxel mesh division to generate a fusion voxel mesh that is independent of the background differential voxel mesh. The fusion voxel mesh is used for subsequent change detection, feature extraction and point cloud fusion operations.

[0106] Furthermore, step S31 includes:

[0107] Step S311: Obtain the functional zoning information of the intelligent automated warehouse. The functional zoning information is generated based on the CAD layout drawing or operational function plan of the intelligent automated warehouse, dividing the global three-dimensional space of the intelligent automated warehouse into multiple non-overlapping functional zones. The functional zones include storage areas, aisle areas, open areas, and equipment operation areas. Each functional zone corresponds to different perception accuracy requirements. Storage areas are used for storing goods and require the highest perception accuracy; aisle areas are used for equipment passage and require medium perception accuracy; open areas and equipment operation areas are used for temporary stacking and equipment operation and require lower perception accuracy.

[0108] Step S312: Establish the mapping relationship between functional zones and fusion voxel resolution. Based on the perception accuracy requirements of different functional zones, assign a corresponding fusion voxel resolution to each functional zone. Functional zones with higher perception accuracy requirements use smaller fusion voxel resolutions. When the fusion voxel resolution is too small, the number of voxels increases dramatically, leading to excessive computation. When the fusion voxel resolution is too large, it cannot meet the accuracy requirements of the corresponding functional zone. For example, the fusion voxel resolution corresponding to the storage area is 2 cm × 2 cm × 2 cm, which can meet the ±1 cm accuracy requirement for pallet docking and storage status detection; the fusion voxel resolution corresponding to the passageway area is 5 cm × 5 cm × 5 cm, which can meet the ±5 cm accuracy requirement for equipment navigation; and the fusion voxel resolution corresponding to the open area and equipment operation area is 20 cm × 20 cm × 20 cm, which can meet the ±10 cm accuracy requirement for obstacle detection.

[0109] Step S313: Generate fusion voxels according to the functional zone to which each spatial location belongs, using the corresponding fusion voxel resolution; assign a unique three-dimensional index to each fusion voxel to establish a mapping relationship between the fusion voxel and the spatial location; the fusion voxel mesh is independent of the background differential voxel mesh used for dynamic and static separation in step S2. The background differential voxel mesh uses a uniform 10 cm × 10 cm × 10 cm resolution and is specifically used for dynamic object detection; the fusion voxel mesh uses an adaptive resolution and is specifically used for high-precision fusion of static backgrounds. The two perform their respective functions and do not interfere with each other.

[0110] Specifically, step S31 is the spatial foundation for change detection and subsequent fusion operations. Its core design addresses the contradiction between the varying perception accuracy requirements of different functional zones in an intelligent automated warehouse and the computational burden of fusion. This is achieved through differentiated voxel grid partitioning, balancing perception accuracy with system real-time performance. In an intelligent automated warehouse, the storage area requires the highest perception accuracy, the aisle area requires medium accuracy, and the open area and equipment operation area require lower accuracy. Using a uniform voxel resolution would either fail to meet the high-precision requirements of the storage area or result in an excessive number of voxels and redundant computation in the open area. Furthermore, the background differential voxel grid used for dynamic and static separation has different functions than the voxel grid used for fusion. Sharing the same grid would cause interference between dynamic detection and static fusion, affecting their effectiveness. Step S31 obtains functional zoning information through S311, dividing functional zones according to the warehouse CAD layout and operational planning, and clarifying the perception requirements of different areas; through S312, it establishes the mapping relationship between functional zones and voxel resolution, assigning appropriate resolutions to different zones to balance accuracy and computational load; through S313, it generates independent fusion voxel meshes, each performing its own function with the background differential voxel mesh to avoid mutual interference, providing accurate spatial units for subsequent change detection and fusion operations. Without step S31, using a uniform voxel resolution will lead to insufficient fusion accuracy in the storage area, excessive computational load in the open area, or mutual interference between the fusion voxel mesh and the background differential voxel mesh, resulting in inaccurate change detection, low fusion efficiency, and inability to adapt to the perception requirements of different functional zones in the warehouse.

[0111] Step S32: Obtain the previous frame static point cloud data set and the previous frame fusion result data; the previous frame static point cloud data set contains discrete point data of all LiDAR determined to be static points in the previous frame, and the previous frame fusion result data contains the previous frame fusion coordinates, previous frame fusion intensity and previous frame voxel number information for each fusion voxel.

[0112] Step S33: For each LiDAR, compare its current frame static point cloud data set with the previous frame static point cloud data set on a voxel-by-voxel basis according to the fused voxel grid generated in step S31, and calculate the change in each fused voxel.

[0113] Furthermore, step S33 includes:

[0114] Step S331: For each lidar, assign all discrete points in its current frame static point cloud data set to the corresponding fusion voxels according to their three-dimensional coordinates;

[0115] Step S332: For each non-empty fusion voxel, calculate the average coordinates and number of points of the static point cloud within the fusion voxel in the current frame, and the average coordinates and number of points of the static point cloud within the fusion voxel in the previous frame. The non-empty fusion voxel refers to a fusion voxel grid that contains at least one LiDAR static point cloud discrete point in the current or previous frame. That is, there is point cloud data within the fusion voxel that can be used for comparative analysis, and the average coordinates, number of points, and change can be calculated. If a fusion voxel has no static point cloud discrete points of any LiDAR in the current and previous frames, it is a blank voxel and does not participate in the change calculation. The average coordinates are the arithmetic mean of the three-dimensional coordinates of all discrete points within the fusion voxel, and the number of points is the total number of discrete points within the fusion voxel.

[0116] Step S333: Calculate the coordinate change based on the distance between the average coordinates of the static point cloud within the fused voxel in the current frame and the average coordinates of the static point cloud within the fused voxel in the previous frame;

[0117] Step S334: Calculate the change in the number of points based on the absolute difference between the number of static points in the fused voxel in the current frame and the number of static points in the fused voxel in the previous frame.

[0118] Specifically, step S33 is the core quantification step for change detection. Its purpose is to accurately calculate the change in each fused voxel by comparing voxel-by-voxel static point clouds of adjacent frames from a single LiDAR scanner. This provides an objective quantitative basis for subsequent voxel labeling of changes, solving the problem of missed changes caused by the inability to capture subtle changes in the static point cloud. In intelligent automated warehouses, changes in the static background are usually quite subtle. Without voxel-by-voxel comparison, it is difficult to capture these subtle changes. At the same time, the point cloud of a single LiDAR scanner exhibits discreteness; the coordinate fluctuations of a single discrete point may be due to measurement errors rather than actual changes. Therefore, it is necessary to weaken the influence of errors from individual points by comparing the average coordinates and the number of points, ensuring the accuracy of the change calculation. Step S33 assigns the point cloud to the corresponding fusion voxels through S331, establishing the correspondence between the point cloud and the fusion voxels; through S332, it calculates the average coordinates and number of points of the non-empty voxels, weakening the measurement error of individual discrete points and improving the stability of the data; through S333 and S334, it calculates the change in coordinates and the change in the number of points, respectively, comprehensively quantifying the changes of voxels from two dimensions: spatial location and point cloud density, ensuring that even subtle changes can be accurately captured. If step S33 is missing, the change in the fusion voxels cannot be accurately calculated, and the subsequent voxel marking in S34 will lack objective basis. This will either fail to capture subtle changes such as the increase or decrease of goods in the storage location, leading to missed changes, or misjudge measurement errors as real changes, leading to misjudgment of voxels and affecting the targeting and accuracy of subsequent fusion operations.

[0119] Step S34: Based on the preset change threshold and the calculated change amount of the fused voxel, mark the changed voxel detected by each lidar;

[0120] The preset change thresholds include preset coordinate change thresholds and preset point number change thresholds, both dynamically set according to the resolution of the corresponding fused voxel. Higher voxel resolution results in stronger detection of minute changes, leading to smaller preset change thresholds; lower voxel resolution results in higher tolerance to noise, leading to larger preset change thresholds. For example, for a fused voxel in a cargo area with a 2cm resolution, the preset coordinate change threshold is set to 1cm, which is half the voxel resolution, effectively distinguishing between real physical changes and point cloud measurement noise. The preset point number change threshold is set to 3 points; when the number of points within a voxel changes by more than 3, it can be determined that the cargo status in that area has changed. For a fused voxel in a passageway area with a 5cm resolution, the preset coordinate change threshold is set to 2.5cm, and the preset point number change threshold is set to 2 points. For a fused voxel in an open area with a 20cm resolution, the preset coordinate change threshold is set to 10cm, and the preset point number change threshold is set to 1 point.

[0121] When the coordinate change of a fused voxel corresponding to a certain lidar is greater than a preset coordinate change threshold, or the change in the number of points is greater than a preset number of points change threshold, the fused voxel is marked as a changed voxel detected by the lidar; otherwise, it is marked as a static voxel detected by the lidar.

[0122] Step S35: Use a multi-radar voting mechanism to merge all the variable voxels detected by the lidar to generate a variable voxel set and a static voxel set;

[0123] The multi-radar voting mechanism is as follows: as long as any lidar detects a change in a fused voxel, the fused voxel is marked as a changed voxel; only when all lidars detect a fused voxel as a static voxel is the fused voxel marked as a static voxel. This mechanism can effectively avoid missed changes caused by single lidars due to viewpoint obstruction, local noise, or sparse point clouds, and ensure the sensitivity of change detection.

[0124] Step S36: Determine the set of voxels to be processed. Use the variable voxel set as the set of voxels to be processed. Perform subsequent feature extraction, weight calculation and fusion operations only on the fused voxels in this set. For all fused voxels in the static voxel set, directly reuse the fusion coordinates and fusion intensity of the previous frame stored in the storage unit without any recalculation.

[0125] Specifically, step S3 is a crucial step in achieving accurate change detection and improving fusion efficiency. Its core design addresses issues such as inaccurate static background change detection in intelligent automated warehouses, the large computational load of full-voxel fusion, and the tendency for single-radar change detection to miss changes. This ensures that subsequent fusion operations only target the changed areas, balancing fusion accuracy and real-time performance. The static background of an intelligent automated warehouse is not entirely static; changes such as the addition or removal of goods in storage locations and slight shifts in shelves need to be accurately detected. If it's impossible to distinguish between changing and static voxels, performing fusion calculations on all voxels would lead to excessive computational load and decreased real-time performance. Furthermore, single-radar systems are prone to missed changes due to viewpoint obstruction and sparse point clouds, necessitating multi-radar collaboration to improve detection sensitivity. Step S3 adapts to the perception needs of different functional zones through differentiated voxel segmentation in S31; acquires the previous frame data in S32 to provide a benchmark for comparison between adjacent frames; calculates voxel changes in S33 to accurately capture subtle changes in the static background; marks single-radar changing voxels in S34, dynamically setting thresholds based on voxel resolution to ensure marking accuracy; uses a multi-radar voting mechanism in S35 to avoid single-radar missed detections and improve the sensitivity of change detection; and determines the set of voxels to be processed in S36, performing subsequent processing only on changing voxels, while reusing the previous frame's results for static voxels, significantly reducing computational load. Without step S3, it is impossible to accurately distinguish between changing and static voxels, either leading to system lag due to full voxel calculation or missed static background changes, resulting in the fused point cloud map not updating in a timely manner and failing to reflect the true state of the warehouse's static background. Furthermore, the missed detection issues of single-radar detection will affect the completeness of the fusion results.

[0126] Step S4: For each variable voxel in the set of voxels to be processed, extract the multi-dimensional quality features of each LiDAR within that variable voxel, and perform differential standardization processing on all extracted quality features to obtain the quality feature scores of each LiDAR within that variable voxel; the multi-dimensional quality features are used to comprehensively quantify the perception accuracy and reliability of each LiDAR within the corresponding voxel, providing an objective basis for subsequent weight calculation.

[0127] Please see Figure 3 Furthermore, step S4 includes:

[0128] Step S41: Based on the variable voxels generated in step S3, combined with the current frame static point cloud data set output in step S2 and the unified global coordinate system in step S14, all discrete points in the current frame static point cloud data set are assigned to the corresponding variable voxels according to the three-dimensional coordinates, and the point cloud mapping relationship between each lidar and each variable voxel is established.

[0129] For each variable voxel, if a certain lidar has no discrete points within that variable voxel, then all quality feature scores of that lidar within that variable voxel are set to 0 and will not participate in the subsequent weight calculation and fusion process.

[0130] Step S42: Based on the point cloud mapping relationship, extract the multi-dimensional quality features of each lidar in each changing voxel. The multi-dimensional quality features include basic geometric features and echo features.

[0131] Furthermore, step S42 includes:

[0132] Step S421: Based on the set of discrete points belonging to the same lidar and the same variable voxel in the point cloud mapping relationship in step S41, calculate the basic geometric features; the basic geometric features are used to reflect the geometric perception accuracy of the lidar in the variable voxel, and are the core indicators for measuring the spatial positioning accuracy of the lidar, including point cloud density, plane fitting error, average ranging distance and average incident angle.

[0133] Furthermore, step S421 includes:

[0134] Step S4211: Calculate the point cloud density. The point cloud density is the ratio of the number of discrete points of the LiDAR within the voxel after preprocessing in step S1 to the volume of the voxel. It is used to reflect the sampling density of the point cloud within the voxel. The higher the point cloud density, the more fully the LiDAR samples the voxel, and the higher the sensing accuracy. The point cloud density ρ is calculated using the formula: ρ = n / V, where n is the number of discrete points of the LiDAR within the voxel, and V is the volume of the voxel. The voxel volume is calculated based on the fusion voxel resolution of its functional partition.

[0135] Step S4212: Calculate the plane fitting error. The plane fitting error is the root mean square error of fitting the discrete points of the LiDAR under the current changing voxel within the point cloud mapping relationship of step S41 to the point cloud plane. It is used to reflect the geometric consistency of the point cloud within the changing voxel. The smaller the plane fitting error, the closer the point cloud is to the ideal plane, and the higher the geometric perception accuracy. For changing voxels containing only a single discrete point, plane fitting cannot be performed. The plane fitting error is set to a preset maximum error value. The preset maximum error value refers to the maximum perception error tolerance value of the corresponding functional area of ​​the intelligent automated warehouse, which is determined by combining the maximum measurement error of the LiDAR and the point cloud noise level. The setting principle is not to exceed 1 / 2 of the resolution of the fused voxel of the corresponding functional area, to ensure that when a single discrete point cannot fit the plane, its error value will not affect the differentiation effect of subsequent quality features. For example, the preset maximum error value is 10 cm.

[0136] Step S4213: Calculate the average ranging distance. The average ranging distance is the arithmetic mean of the distances from all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship in step S41 to the origin of the lidar's own coordinate system. It is used to reflect the spatial distance between the changing voxel and the lidar. The closer the distance, the higher the ranging accuracy of the lidar and the stronger the reliability of geometric perception.

[0137] Step S4214: Calculate the average incident angle. The average incident angle is the arithmetic mean of the angles between the laser incident direction and the normal vector of all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship in step S41. It is used to reflect the observation angle of the lidar on the surface of the changing voxel. The smaller the incident angle, the stronger the laser reflection signal and the better the stability of geometric perception. The normal vector is obtained by analyzing and calculating the point cloud within the voxel using the principal component analysis method. For voxels containing only a single discrete point, the normal vector is set to the default value (0,0,1) to ensure the completeness of feature extraction.

[0138] Step S422: Based on the set of discrete points belonging to the same lidar and the same variable voxel in the point cloud mapping relationship in step S41, extract the echo features; the echo features are used to reflect the quality of the lidar echo signal in the variable voxel, and are the core indicators for measuring the reliability of lidar signal reception, including average echo intensity, number of echoes and echo interval.

[0139] Furthermore, step S422 includes:

[0140] Step S4221: Calculate the average echo intensity. The average echo intensity is the arithmetic mean of the echo intensity of all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship in step S41. It is used to reflect the reflection characteristics of the object surface corresponding to the changing voxel. The closer the echo intensity is to the preset optimal value, the better the quality of the echo signal retained after the preprocessing in step S1, and the higher the reliability of the lidar in sensing the changing voxel.

[0141] Step S4222: Calculate the number of echoes. The number of echoes is the maximum value of the echo numbers of all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship in step S41. It is used to reflect whether there are semi-transparent objects, multi-layer structures or obstructions in the changing voxel. The more echoes there are, the better the penetration of the lidar to the changing voxel and the more comprehensive the perception.

[0142] Step S4223: Calculate the echo interval, which is the arithmetic mean of the time differences between adjacent echoes of all multi-echo points of the lidar under the current changing voxel within the point cloud mapping relationship in step S41. It is used to reflect the spacing of the multi-layer structure within the changing voxel and indirectly reflect the thickness or number of stacked layers of the object. For voxels without multi-echo points, the echo interval is set to 0 to ensure the integrity of feature extraction.

[0143] Specifically, step S42 is the core step of quality feature extraction. Its design aims to comprehensively and accurately quantify the perception accuracy and reliability of each LiDAR within the variable voxel, providing a multi-dimensional objective basis for subsequent weight calculation and solving the problem that a single feature cannot comprehensively measure the perception quality of the LiDAR. In intelligent automated warehouses, different LiDARs have different deployment locations and observation angles, resulting in differences in perception quality within the same variable voxel. These differences are reflected in both geometric positioning accuracy and echo signal quality. Relying on a single feature alone cannot comprehensively and accurately reflect these differences; if feature extraction is incomplete, it will lead to deviations in subsequent weight calculations, failing to highlight the contribution of high-reliability radar data. Step S42 extracts basic geometric features from S421, including point cloud density, plane fitting error, average ranging distance, and average incident angle. From a spatial positioning perspective, this quantifies the geometric perception accuracy of the LiDAR. Point cloud density reflects sampling sufficiency, plane fitting error reflects geometric consistency, average ranging distance reflects ranging accuracy, and average incident angle reflects observation stability. Step S422 extracts echo features, including average echo intensity, echo quantity, and echo interval. From a signal reception perspective, this quantifies the signal reliability of the LiDAR. Average echo intensity reflects signal quality, echo quantity reflects perception comprehensiveness, and echo interval reflects object structure perception capability. Simultaneously, reasonable default values ​​are set for special cases to ensure the completeness of feature extraction. Without step S42, multi-dimensional quality features cannot be obtained, and subsequent weight calculations will lack objective basis, relying solely on subjective experience to allocate weights. This leads to unreasonable weight allocation, failing to accurately reflect the differences in perception reliability among various LiDARs, and affecting the accuracy of the fusion results.

[0144] Step S43: All extracted quality features are divided into three categories—positive features, negative features, and moderate features—based on their physical meaning. Differential standardization processing is then performed to uniformly map quality features with different dimensions and value ranges to the [0,1] interval, eliminating the influence of dimensional differences on subsequent weight calculations. This yields the quality feature scores for each lidar within the variable voxel, ensuring that the standardization results accurately reflect the quality. The differential standardization processing includes positive feature standardization, negative feature standardization, and moderate feature standardization.

[0145] Furthermore, step S43 includes:

[0146] Step S431: Positive feature standardization, where the positive feature is the feature whose larger value corresponds to better perception quality, including point cloud density and echo number; the original values ​​of this type of feature in the same variable voxel of all lidars are normalized to their maximum and minimum values, with the minimum value of the feature in all lidars as the lower limit and the maximum value as the upper limit, and each original feature value is mapped to between 0 and 1. The larger the original value, the higher the score after standardization.

[0147] Step S432: Negative feature standardization. The negative features are those with smaller values ​​corresponding to better perception quality, including plane fitting error, average ranging distance, average incident angle, and echo interval. The original values ​​of this type of feature in all lidars within the same variable voxel are subjected to inverse maximum and minimum value normalization. Similarly, the minimum and maximum values ​​of this feature in all lidars are used as the upper and lower limits. The smaller the original value, the higher the score after standardization.

[0148] Step S433: Moderate feature standardization. The moderate feature is one whose value is closer to the preset optimal value, corresponding to better perception quality, including average echo intensity. Using the preset optimal value of this feature as the central benchmark, the deviation between the original value and the optimal value of this feature for each LiDAR is compared. The smaller the deviation, the higher the score after standardization. Finally, it is uniformly mapped to the range of 0 to 1. The optimal value of the average echo intensity is set according to the technical parameters of the LiDAR and the object reflection characteristics of the smart warehousing scenario. For example, the optimal value is 125. The original values ​​of the average echo intensity of LiDARs A and B are respectively... Taking 115 and 140 as examples, the absolute deviation of radar A from the optimal value is calculated to be 10, and the absolute deviation of radar B from the optimal value is 15. Taking the maximum deviation of 15 in this region as the normalization benchmark, the standardized score of radar A is (15-10) / 15≈0.33, and the standardized score of radar B is (15-15) / 15=0. This completes the calculation of the single feature quality score. The range of the quality feature score obtained after standardization is [0,1]. The higher the score, the better the quality of the lidar on this feature, which provides a standardized input basis for the subsequent weight calculation of the combined entropy weight method.

[0149] Specifically, step S4 is a fundamental prerequisite for weight calculation. Its core design addresses the issues of inconsistent dimensions and wide ranges of values ​​for multi-dimensional quality features, as well as the inability of a single feature to comprehensively measure the reliability of LiDAR perception. Through feature extraction and differentiated standardization, it provides standardized and comprehensive input for subsequent weight construction. In intelligent automated warehouses, the dimensions and value ranges of different quality features vary significantly. Without standardization, they cannot be directly used for weight calculation. Furthermore, different features have different impacts on perception reliability; using a uniform standardization method would result in feature scores failing to accurately reflect the quality of perception. Step S4 establishes a point cloud mapping relationship through S41 to ensure the targeted nature of feature extraction, setting the feature score to 0 for radars without point clouds to avoid interference from invalid data. Through S42, it extracts multi-dimensional quality features, comprehensively covering both geometric and echo dimensions, fully quantifying perception reliability. Through differentiated standardization in S43, it uses corresponding standardization methods based on feature type to uniformly map all features to the [0,1] interval, eliminating dimensional differences and ensuring that feature scores accurately reflect the quality of perception. If step S4 is missing, multi-dimensional features cannot be uniformly compared and calculated, the weight calculation will lose its foundation, and subsequent combination weight construction and comprehensive fusion weight calculation cannot be carried out, or the weight allocation will be unreasonable due to inaccurate feature scores, ultimately affecting the accuracy and reliability of the fusion result.

[0150] Step S5: Based on the functional partition to which the variable voxel belongs, construct a combined weight by combining prior weights and entropy weights. Calculate the comprehensive score of each lidar based on the quality feature score. After normalization, obtain the comprehensive fusion weight of each lidar within the corresponding variable voxel. The comprehensive fusion weight is used to characterize the perception reliability ratio of each lidar within the current voxel, providing a quantitative basis for subsequent multi-lidar point cloud weighted fusion.

[0151] Please see Figure 4 Furthermore, step S5 includes:

[0152] Step S51: Obtain the prior weights of the quality features corresponding to the functional zones. The prior weights refer to fixed weight values ​​pre-assigned to each quality feature based on engineering experience and historical measured data in the field of intelligent warehousing. These weights are set according to the operational attributes, perception requirements, and LiDAR perception patterns of different functional zones in the intelligent automated warehouse. The storage area has the highest requirements for positioning accuracy, so the weight of geometric quality features is higher. The aisle area and open area emphasize signal stability, so the weight of echo quality features is correspondingly increased. For example, the prior weights of point cloud density and plane fitting error are relatively higher in the storage area, while the prior weights of average echo intensity and average ranging distance are relatively higher in the aisle area. The prior weights for the storage area can be: point cloud density 0.25, plane fitting error 0.25, average distance measurement 0.15, average angle of incidence 0.15, average echo intensity 0.10, echo quantity 0.05, and echo interval 0.05; the prior weights for the passage area are: point cloud density 0.15, plane fitting error 0.15, average distance measurement 0.20, average angle of incidence 0.10, average echo intensity 0.20, echo quantity 0.10, and echo interval 0.10; the prior weights for the open area are: point cloud density 0.10, plane fitting error 0.10, average distance measurement 0.20, average angle of incidence 0.10, average echo intensity 0.25, echo quantity 0.15, and echo interval 0.10. All the above prior weights are statistically calibrated based on measured data from warehousing scenarios, and the sum of all prior weights for quality characteristics within each functional area is 1.

[0153] Step S52: Based on the quality feature scores of all LiDARs output in Step S4 within the current changing voxel, quantify the data dispersion and information contribution of each quality feature, and calculate the objective entropy weight of each quality feature. Data dispersion refers to the difference in scores for the same quality feature among different LiDARs; the greater the difference, the higher the dispersion, indicating that the feature can more clearly distinguish the perception quality differences between different LiDARs. Information contribution refers to the effective information content of the quality feature in judging the perception reliability of the LiDAR; the higher the dispersion, the greater the information contribution, and the greater the corresponding objective entropy weight. By calculating the information entropy of each feature score, interference from quality features with high information redundancy and weak distinguishing ability is eliminated, resulting in objective weights driven solely by the data itself, avoiding subjective bias caused by prior human experience.

[0154] Furthermore, step S52 includes:

[0155] Step S521: Construct a standardized score matrix for all LiDAR quality features under the current changing voxel. The rows of the standardized score matrix correspond to different LiDARs, the columns correspond to each quality feature, and the matrix elements are the corresponding quality feature scores output in step S4. This matrix clearly presents the score distribution of the same quality feature among different LiDARs, providing a basis for subsequent quantification of data dispersion and calculation of information contribution.

[0156] Step S522: Calculate the proportion of each quality feature and its information entropy under all lidars in turn. Calculate the redundancy coefficient based on the information entropy difference. Finally, normalize to obtain the objective entropy weight of each quality feature. The smaller the information entropy, the stronger the distinguishability of the feature among different lidars, and the larger the corresponding objective entropy weight. The larger the information entropy, the more homogeneous the feature data is and the less effective information it contains, and the smaller the corresponding objective entropy weight.

[0157] The weight of the quality feature refers to the ratio of the standardized feature score of a single lidar to the sum of the standardized scores of all lidars for the same quality feature. It is used to characterize the relative proportion of a single lidar in this feature dimension. The larger the difference in the ratio, the higher the data dispersion of this feature.

[0158] The information entropy refers to an index of information disorder calculated based on the proportion of quality characteristics of each lidar. It measures the degree of discrete difference of this quality characteristic among multiple lidars and directly reflects the information contribution of this characteristic in distinguishing the perception quality of different lidars. Information entropy is negatively correlated with information contribution. The specific calculation formula is as follows:

[0159]

[0160] in, For the first Information entropy of a quality feature The number of lidar units involved in the calculation. For the first The lidar in the first The proportion of quality characteristics under each quality characteristic, when At that time, take To avoid the logarithm becoming meaningless;

[0161] The specific calculation logic of objective entropy weight is as follows: First, the information utility value is calculated based on the information entropy of each quality feature. The information utility value is the difference between the preset maximum information entropy and the current feature information entropy. The smaller the information entropy, the larger the information utility value. Then, the information utility value of each quality feature is globally normalized. The weight value obtained after normalization is the objective entropy weight of the corresponding quality feature. The larger the information utility value, the stronger the feature's distinguishing ability and the higher its information contribution, and the larger the final objective entropy weight. Among them, the preset maximum information entropy is determined according to the number of lidars involved in the calculation to ensure that the calculation of information utility value has a unified benchmark. The global normalization process adopts the extreme value normalization method, mapping the information utility values ​​of all quality features to the [0,1] interval, and the sum of the objective entropy weights of all quality features is 1.

[0162] Specifically, step S52 is the core step in calculating objective entropy weights. Its design aims to quantify the data dispersion and information contribution of each quality feature to obtain objective weights driven by the data itself, thus addressing the problem of subjective bias and inability to adapt to real-time point cloud data changes when relying solely on prior weights. In intelligent automated warehouses, the distinguishing ability of each quality feature varies across different scenarios. Some features show significant score differences across different radars, effectively distinguishing perceived quality, while others exhibit highly homogeneous scores and low information contribution. Relying solely on prior weights cannot adapt to such real-time changes, leading to weight allocations that do not match the actual perceived quality. Step S52 constructs a standardized score matrix through S521, clearly presenting the score distribution of the same feature across different radars, providing a foundation for quantifying dispersion. Through S522, it calculates the quality feature proportion, information entropy, and objective entropy weights. The smaller the information entropy, the stronger the feature's distinguishing ability and the higher its information contribution, resulting in a larger objective entropy weight, and vice versa. This method eliminates the interference of redundant features, highlighting the role of effective features and obtaining objective weights that are consistent with real-time data. If step S52 is missing, the weight calculation will rely solely on prior experience, which cannot adapt to changes in real-time point cloud data. Redundant features will interfere with weight allocation, causing the objective entropy weight to fail to reflect the inherent patterns of the data. The objectivity and adaptability of the combined weights will be greatly reduced, and the real-time perception reliability of each LiDAR will not be accurately reflected.

[0163] Step S53: The prior weights obtained in step S51 and the objective entropy weights obtained in step S52 are weighted and fused according to a preset weight ratio to obtain the combined weights of each quality feature. The preset weight ratio is set according to the degree of disturbance in warehouse operations. When warehouse goods frequently enter and exit and the operation is dynamic, the proportion of objective entropy weights is increased to adapt to real-time changes. When warehouse goods are stationary for a long time and the operating environment is stable, the proportion of prior weights is increased to conform to the inherent laws of the scenario. For example, the ratio of prior weights to objective entropy weights is 4:6, which takes into account both the prior perception laws in the warehousing scenario and makes full use of the objective information of real-time point cloud data, so that the weight allocation takes into account both empirical rationality and real-time adaptability.

[0164] Step S54: For each LiDAR, multiply its quality feature scores within the current changing voxel by the combined weights of the corresponding quality features in sequence, and then sum them up to obtain the comprehensive score of the LiDAR; the higher the comprehensive score, the stronger the overall perception reliability of the LiDAR within the current changing voxel.

[0165] Step S55: Perform global normalization on the comprehensive scores of each LiDAR to obtain the comprehensive fusion weight, so that the sum of the comprehensive fusion weights of all LiDARs participating in the fusion is 1; the normalized comprehensive fusion weight can be directly used for subsequent weighted fusion calculation. The higher the weight ratio, the greater the contribution of the LiDAR's perception data to the current voxel fusion result.

[0166] Step S6: Based on the comprehensive fusion weights obtained in Step S5, perform incremental voxel-level weighted fusion operation on the changing voxels, and combine the fusion results of the previous frame of static voxels to generate the fused point cloud map of the current frame; the fused point cloud map integrates the effective perception data of all LiDARs, has higher geometric accuracy and signal reliability, and provides a unified three-dimensional perception foundation for the upper-level application of intelligent automated warehouses.

[0167] Specifically, step S5 is the core step in achieving weighted fusion of multi-radar point clouds. Its design focuses on addressing the issue that a single weight cannot simultaneously account for both the experience of intelligent automated warehouse scenarios and real-time data changes. By constructing combined weights, accurate comprehensive fusion weights are obtained, providing a quantitative basis for subsequent fusion operations and ensuring that the fusion result highlights the contribution of highly reliable radar data. In intelligent automated warehouses, different functional zones have different perception requirements. Prior weights can align with the inherent rules of the scenario, such as emphasizing geometric accuracy in the storage area and signal stability in the aisle area, but they have subjective biases. Objective entropy weights can adapt to real-time data changes and reflect the inherent patterns of the data, but they lack scenario experience support. Using only a single weight would lead to unreasonable weight allocation, affecting fusion accuracy. Step S5 obtains prior weights through S51, aligning with the perception needs of different functional zones and relying on scenario experience and historical data to ensure the rationality of the weights; through S52, it calculates objective entropy weights based on real-time point cloud data to ensure the objectivity and adaptability of the weights; through S53, it fuses the prior weights and objective entropy weights, setting the ratio according to the degree of disturbance in warehouse operations, taking into account both scenario experience and real-time changes; through S54, it calculates the comprehensive score to comprehensively measure the perception reliability of each radar; and through S55, it performs normalization processing to obtain a comprehensive fusion weight that can be directly used for fusion calculation. If step S5 is missing, a reasonable comprehensive fusion weight cannot be obtained, and subsequent fusion operations will be unable to quantify the perception contribution of each radar, and can only use equal weights or subjective weights for fusion, resulting in the inability to reflect the advantages of high-reliability radar data, and the inability to guarantee the accuracy and reliability of the fusion results.

[0168] Furthermore, step S6 includes:

[0169] Step S61: For each variable voxel, extract the static point cloud data of each participating LiDAR within the variable voxel from the point cloud mapping relationship established in step S41, including the three-dimensional coordinates and echo intensity information of each discrete point. At the same time, retrieve the comprehensive fusion weights of each LiDAR output in step S55 to ensure the integrity and accuracy of the data required for the fusion calculation. For LiDARs with a comprehensive fusion weight of 0, the point cloud data within their corresponding voxel will not participate in this fusion calculation.

[0170] Step S62: For each variable voxel, the three-dimensional coordinates and echo intensity are weighted and fused separately based on the comprehensive fusion weight to obtain the fused coordinates and fused intensity of the variable voxel. The fusion process strictly follows the contribution ratio of the comprehensive fusion weight to ensure that the lidar data with higher perception reliability dominates the fusion result.

[0171] Furthermore, step S62 includes:

[0172] Step S621: Calculate the average 3D coordinates of all static point clouds within the current changing voxel for each participating LiDAR. This average 3D coordinate is the arithmetic mean of the 3D coordinates of all discrete points within that voxel for the corresponding LiDAR, effectively reducing the impact of measurement errors at individual discrete points. Subsequently, using the comprehensive fusion weight of each LiDAR as the weight, the average 3D coordinates of all participating LiDARs are weighted and summed to obtain the final fused coordinates of the changing voxel. The fused coordinates take into account the spatial positioning advantages of multiple LiDARs, improving the positioning accuracy of the voxel's spatial position. For example, let the first... The integrated fusion weight of the LiDAR is Its average three-dimensional coordinates are Then the final fused coordinates The calculation formula is:

[0173]

[0174] The sum of the weights of all participating radar fusion systems is 1.

[0175] Step S622: Using the same weighted fusion logic as the fusion coordinates, calculate the average echo intensity of all static point clouds within the current changing voxel for each participating lidar. This average echo intensity is the arithmetic mean of the echo intensities of all discrete points within the changing voxel for the corresponding lidar, which can weaken the interference of individual discrete point echo noise. Subsequently, using the comprehensive fusion weight of each lidar as the weight, the average echo intensities of all participating lidars are weighted and summed to obtain the final fusion intensity of the changing voxel. The fusion intensity can more accurately reflect the reflection characteristics of the object surface corresponding to the voxel. For example, let the first... The average echo intensity of the lidar is The final fusion strength The calculation formula is:

[0176]

[0177] Specifically, step S62 is the core calculation step of variable voxel weighted fusion. Its design aims to fully leverage the sensing advantages of each LiDAR based on the comprehensive fusion weight, solving the problems of measurement error and noise interference and low fusion accuracy when directly fusing discrete points, and ensuring that the fusion result of variable voxels has high geometric accuracy and signal reliability. In intelligent automated warehouses, the sensing reliability of different LiDARs within the same variable voxel varies. The comprehensive fusion weight can quantify this difference. If discrete points are directly weighted for fusion, the measurement error and noise of individual discrete points will affect the fusion result. At the same time, the fusion result needs to take into account both spatial position and signal characteristics, and the fusion logic of both needs to be consistent to ensure the integrity of the fusion result. Step S62 calculates the average three-dimensional coordinates of each LiDAR through S621, weakening the measurement error of individual discrete points, and then performs weighted summation based on the comprehensive fusion weight to highlight the positioning advantage of high-weight LiDARs and improve the accuracy of fused coordinates. Through S622, the same logic is used to calculate the average echo intensity and perform weighted fusion, weakening the interference of echo noise, accurately reflecting the reflection characteristics of the object surface, and ensuring the reliability of the fused intensity. Without step S62, accurate weighted fusion of variable voxels cannot be achieved. Either the discrete points are directly fused, leading to amplified errors, or the perception differences of each radar cannot be reflected, resulting in insufficient accuracy of fused coordinates and fused intensity, which cannot meet the perception requirements of upper-level warehouse applications.

[0178] Step S63: Combine the fusion coordinates and fusion intensity of all the changed voxels with the fusion coordinates and fusion intensity of the static voxels determined in step S36 from the previous frame. Static voxels directly reuse the fusion results from the previous frame without repeated calculations, which ensures the continuity of the fused point cloud map, significantly reduces the computational load of the system, and improves the real-time performance of the fusion processing. After integration, a complete fused point cloud map of the current frame is obtained. This map contains the fusion coordinates and fusion intensity information of all non-empty fused voxels globally, and each voxel retains its functional partition identifier and 3D index information.

[0179] Specifically, step S6 is the final output stage of the entire multi-radar point cloud weighted fusion method. Its core design addresses the problems of high computational cost and poor real-time performance in full-voxel fusion, as well as the inability to effectively integrate multi-radar data. Through incremental voxel-level weighted fusion, a high-precision, high-reliability fused point cloud map is generated, providing a unified 3D perception foundation for upper-level applications in intelligent automated warehouses. In intelligent automated warehouses, static voxels constitute a large proportion and change little. Re-fusion of all voxels would lead to excessive computational cost and decreased real-time performance. Furthermore, the advantage of multi-radar data lies in their complementarity; if they cannot be effectively integrated, the fusion result will fail to take into account the perceptual advantages of each radar, resulting in insufficient accuracy and reliability. Step S6 extracts the data required for fusion through S61, ensuring data integrity and accuracy, and removing invalid data with a weight of 0. Step S62 performs weighted fusion of the changing voxels, fully leveraging the sensing advantages of each radar to improve fusion accuracy. Step S63 stitches the fusion results of changing and static voxels together, with static voxels reusing the results from the previous frame. This ensures the continuity of the fused map while significantly reducing computational load and improving system real-time performance. The final generated fused point cloud map integrates the effective sensing data from all LiDARs, possessing high geometric accuracy and signal reliability. Without step S6, the point cloud data from multiple radars cannot be effectively integrated, a unified fused point cloud map cannot be generated, or the system may fail to output results in real time due to excessive computational load, thus failing to provide reliable 3D sensing support for upper-level applications such as location detection, equipment navigation, and security monitoring in intelligent automated warehouses.

[0180] Example 2

[0181] This embodiment introduces a multi-radar point cloud weighted fusion system based on the entropy weighting method, including:

[0182] Data Acquisition Module: Responding to the acquisition trigger signal of the intelligent automated warehouse, this module synchronously acquires the current frame point cloud data from all LiDARs. The acquisition trigger signal can be automatically generated according to a preset timing period or manually triggered by the intelligent automated warehouse control system. The preset timing period can be adjusted according to the dynamic operation level of the warehouse and the global perception accuracy requirements. This module is also used to independently perform preprocessing operations on the raw point cloud data acquired by each LiDAR. Specifically, this includes removing distance and intensity anomalies based on the 3D coordinates and echo intensity of discrete points, removing outlier noise points using statistical filtering methods, and performing motion distortion correction on LiDARs installed on mobile devices. Subsequently, all preprocessed point cloud data are uniformly converted to the global coordinate system of the intelligent automated warehouse. Finally, the preprocessed point cloud data from all LiDARs are integrated to generate the current frame point cloud data set, providing a unified input for subsequent processing steps.

[0183] The static / dynamic separation module performs independent temporal dynamic / static separation operations on the current frame point cloud data set for a single LiDAR, distinguishing dynamic points from static points and generating a static point cloud data set. Its specific functions include initializing background difference voxels and background difference counters. Background difference voxels are three-dimensional spatial cube units used for temporal background difference. Independent background difference counters need to be maintained for each background difference voxel and each LiDAR to count the number of consecutive point cloud frames appearing within the corresponding voxel for the corresponding LiDAR. The module divides the background difference voxel grid and initializes the counter values. It traverses the discrete points of each LiDAR to determine its corresponding background difference voxel and updates the corresponding counter values. Based on the counter values ​​and a preset counting threshold, it determines the static or dynamic attributes of discrete points within each background difference voxel. Finally, it separates dynamic and static points, merges all dynamic points to generate the current frame dynamic point dataset for individual tracking, and integrates all static points to generate the current frame static point data set, which serves as input for subsequent change detection steps.

[0184] The change detection module performs differentiated fusion voxel mesh division based on the functional zones of the intelligent automated warehouse, and performs independent adjacent frame change detection on the current frame static point cloud data set by a single LiDAR, distinguishing between changing voxels and static voxels. Its specific functions include: acquiring the functional zone information of the intelligent automated warehouse; establishing a mapping relationship between functional zones and fusion voxel resolution; generating fusion voxels with corresponding resolutions according to the functional zone to which each spatial location belongs and assigning a unique 3D index; constructing a fusion voxel mesh independent of the background differential voxel mesh; acquiring the previous frame static point cloud data set and the previous frame fusion result data; comparing the current frame and the previous frame static point cloud data for each LiDAR according to the fusion voxel mesh voxel for each voxel, calculating the coordinate change and point number change of each fusion voxel; marking the changing voxels and static voxels detected by each LiDAR according to a preset change threshold; merging the detection results of all LiDARs using a multi-LiDAR voting mechanism to generate a changing voxel set and a static voxel set; using the changing voxel set as the voxel set to be processed, while the static voxels directly reuse their previous frame fusion result.

[0185] Feature extraction module: This module extracts multi-dimensional quality features of each LiDAR within each variable voxel in the set of voxels to be processed. It then performs differential standardization on all extracted quality features to obtain quality feature scores for each LiDAR within that variable voxel. Specific functions include: assigning discrete points to corresponding variable voxels based on the variable voxels and the current frame's static point cloud data set, establishing a point cloud mapping relationship between each LiDAR and each variable voxel, and setting the corresponding quality feature score to 0 for LiDARs without discrete points; extracting basic geometric features and echo features based on the point cloud mapping relationship. Basic geometric features include point cloud density, plane fitting error, average ranging distance, and average incident angle; echo features include average echo intensity, echo quantity, and echo interval. Based on the physical meaning of the quality features, they are divided into three categories: positive features, negative features, and moderate features. Corresponding standardization methods are used to uniformly map all quality features to the [0,1] interval, eliminating dimensional differences and providing an objective basis for subsequent weight calculations.

[0186] The weight calculation module, based on the functional zone to which the changing voxel belongs, constructs a combined weight using prior weights and entropy weights. It calculates the comprehensive score of each LiDAR based on quality feature scores, and after normalization, obtains the comprehensive fusion weight of each LiDAR within the corresponding changing voxel. Its specific functions include: obtaining the prior weights of quality features corresponding to different functional zones, which are determined based on engineering experience and historical measured data in the field of intelligent warehousing, and allocating different weight proportions according to the perception requirements of the functional zones; constructing a standardized score matrix based on quality feature scores, calculating the quality feature proportion, information entropy, and information utility value of each quality feature, and obtaining the objective entropy weight of each quality feature after normalization; weighting and fusing the prior weights and objective entropy weights according to a preset weight ratio to obtain the combined weight; calculating the comprehensive score of each LiDAR and performing global normalization to obtain the comprehensive fusion weight. This weight is used to characterize the perception reliability proportion of each LiDAR within the current changing voxel, providing a quantitative basis for subsequent weighted fusion.

[0187] Point cloud map generation module: Based on comprehensive fusion weights, it performs incremental voxel-level weighted fusion operations on changing voxels, and combines the fusion results of the previous frame of static voxels to generate the fused point cloud map of the current frame. Its specific functions include: for each changing voxel, extracting the static point cloud data and corresponding comprehensive fusion weights of each participating LiDAR within that voxel, excluding LiDAR data with a comprehensive fusion weight of 0; based on the comprehensive fusion weights, weighted summing of the average 3D coordinates and average echo intensity of the static point cloud of each LiDAR within that changing voxel to obtain the fusion coordinates and fusion intensity of that changing voxel; and stitching together the fusion results of all changing voxels with the fusion results of the previous frame of static voxels to generate a complete fused point cloud map of the current frame. This map contains the fusion coordinates, fusion intensity, functional partition identifiers, and 3D index information of all non-empty fusion voxels globally, providing a unified 3D perception foundation for upper-level applications of intelligent automated warehouses.

[0188] Working principle and its effects:

[0189] This invention addresses the practical needs of multi-radar point cloud fusion in intelligent automated warehouses, enabling synchronous collection, purification, filtering, feature quantification, scientific weighting, and incremental fusion of point cloud data. Ultimately, it generates a high-precision, high-real-time fused point cloud map, solving the pain points of existing fusion technologies while adapting to the differentiated perception needs of warehousing scenarios.

[0190] During operation, the system first responds to the data acquisition trigger signal of the intelligent automated warehouse, synchronously acquiring the current frame point cloud data of all LiDARs based on a unified time reference. Through preprocessing operations such as outlier removal, filtering and noise reduction, motion compensation, and coordinate system unification, the synchronization, accuracy, and uniformity of the point cloud data are ensured, laying a high-quality data foundation for subsequent fusion work and effectively avoiding the decrease in fusion accuracy caused by data distortion, temporal misalignment, or coordinate system inconsistency. Subsequently, through independent temporal dynamic and static separation of each LiDAR, the temporal distribution of the point cloud is statistically analyzed using background differential voxels and counters to accurately distinguish dynamic and static points, eliminating interference from dynamic objects in warehousing operations and obtaining pure static point cloud data, reducing the negative impact of dynamic interference on static background fusion. Based on the functional zoning of the intelligent automated warehouse, differentiated fusion voxel mesh division is performed to adapt to the perception accuracy requirements of different areas. The LiDAR independently calculates changes by comparing voxels in adjacent frames, and uses a multi-radar voting mechanism to select the changing voxels as the objects to be processed. This ensures the sensitivity of change detection while avoiding redundancy caused by full voxel calculation, thus improving real-time processing. For the changing voxels, multi-dimensional geometric and echo quality features are extracted and differentiated and standardized to comprehensively quantify the perception quality of each LiDAR. Then, a combined weight is constructed by combining functional partition prior weights and entropy weighting, taking into account scene experience and real-time data patterns. A reasonable comprehensive fusion weight is obtained through comprehensive score calculation and normalization to ensure that the weight allocation is consistent with the actual perception quality. Finally, incremental voxel-level weighted fusion is used to fuse the coordinates and echo intensity of the changing voxels based on the weights. Static voxels reuse the results from the previous frame, which not only highlights the contribution of high-reliability radar data and improves fusion accuracy, but also significantly reduces the amount of computation, further improving the real-time performance of the system.

[0191] In summary, this invention, through the coordinated operation of each step, forms a complete fusion system from data acquisition to map generation. It effectively solves the problems of low accuracy, poor real-time performance, unscientific weight allocation, and susceptibility to dynamic interference in existing multi-radar point cloud fusion systems. The generated fused point cloud map can accurately reflect the static background state of the entire intelligent automated warehouse, adapt to the perception needs of different functional zones, and provide reliable three-dimensional perception support for upper-level applications such as warehouse location identification, equipment navigation, and safety monitoring, significantly improving the automation and intelligent operation level of intelligent automated warehouses.

[0192] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A weighted fusion method for multi-radar point clouds based on entropy weighting, characterized in that, include: In response to the acquisition trigger signal of the intelligent automated warehouse, the point cloud data of all LiDARs in the current frame are acquired synchronously, point cloud data preprocessing is performed, and point cloud data set of the current frame is generated. Perform independent temporal dynamic and static separation on the current frame point cloud data set using a single LiDAR, distinguish between dynamic and static points, and generate a static point cloud data set; Based on the functional partitioning of the fusion voxel mesh, the current frame static point cloud data set is subjected to single-liDAR independent adjacent frame change detection to distinguish changing voxels from static voxels and generate a set of voxels to be processed. For each variable voxel in the set of voxels to be processed, the multi-dimensional quality features of each lidar within the variable voxel are extracted and differentiated and standardized to obtain the quality feature scores of each lidar within the variable voxel. Based on the functional partition of the variable voxel, a combined weight is constructed by combining prior weight and entropy weight method. The comprehensive score of each lidar is calculated based on the quality feature score. After normalization, the comprehensive fusion weight of each lidar in the corresponding variable voxel is obtained. Based on the comprehensive fusion weight, an incremental voxel-level weighted fusion operation is performed on the changing voxels, and the fused point cloud map of the current frame is generated by combining the fusion result of the static voxels from the previous frame.

2. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 1, characterized in that, The steps for generating the static point cloud dataset include: Initialize the background difference voxel and the background difference counter; the background difference counter is used to count the number of frames in which the point cloud appears consecutively in the corresponding background difference voxel of the corresponding lidar; For each lidar, iterate through all discrete points in the current frame point cloud data set, determine the background difference voxel to which the discrete point belongs based on its three-dimensional coordinates, and update the value of the corresponding background difference counter. Based on the value of the background difference counter and the preset counting threshold, the discrete points within each background difference voxel are determined to be static or dynamic points. All discrete points identified as static points by LiDAR are integrated to generate a static point data set for the current frame.

3. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 1, characterized in that, The step of generating the set of voxels to be processed includes: Based on the functional zoning of the intelligent automated warehouse, differentiated voxel grid division is carried out; Obtain the fusion result data of the previous frame's static point cloud data set and the previous frame's fusion result data; For each LiDAR, its current frame static point cloud data set is compared with the previous frame static point cloud data set according to the fused voxel grid, and the change of each fused voxel is calculated. Based on the preset change threshold and the calculated change amount of the fused voxel, mark the changed voxel detected by each lidar; A multi-radar voting mechanism is used to merge all the variable voxels detected by lidar, generating a variable voxel set and a static voxel set, and the variable voxel set is used as the voxel set to be processed.

4. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 3, characterized in that, The change in each fused voxel includes the change in coordinates and the change in the number of points; The step of calculating the change in each fusion voxel includes: For each LiDAR, all discrete points in its current frame static point cloud data set are assigned to the corresponding fusion voxels according to their three-dimensional coordinates; For each non-empty fused voxel, calculate the average coordinates and number of points of the static point cloud within the fused voxel in the current frame, and the average coordinates and number of points of the static point cloud within the fused voxel in the previous frame. The coordinate change is calculated based on the distance between the average coordinates of the static point cloud within the fused voxel in the current frame and the average coordinates of the static point cloud within the fused voxel in the previous frame. The change in the number of points is calculated based on the absolute difference between the number of static points in the fused voxel in the current frame and the number of static points in the fused voxel in the previous frame.

5. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 1, characterized in that, The steps of extracting multi-dimensional quality features of each lidar within a variable voxel and performing differential standardization to obtain the quality feature scores of each lidar within that variable voxel include: All discrete points in the current frame static point cloud data set are assigned to corresponding variable voxels according to their three-dimensional coordinates, and a point cloud mapping relationship between each lidar and each variable voxel is established. Based on the point cloud mapping relationship, multi-dimensional quality features of each lidar in each changing voxel are extracted. The multi-dimensional quality features include basic geometric features and echo features. All extracted quality features are classified into positive features, negative features, and moderate features according to their physical meaning. Differential standardization processing is performed, and they are uniformly mapped to the [0,1] interval to obtain the quality feature scores of each LiDAR within the variable voxel.

6. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 5, characterized in that, The basic geometric features are used to reflect the geometric perception accuracy of the lidar within the variable voxel, including point cloud density, plane fitting error, average ranging distance, and average incident angle. The point cloud density is the ratio of the number of discrete points of the lidar within the variable voxel after preprocessing to the volume of the variable voxel. The plane fitting error is the root mean square error of fitting the discrete points of the lidar under the current changing voxel within the point cloud mapping relationship to the point cloud plane. The average ranging distance is the arithmetic mean of the distances from all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship to the origin of the lidar's own coordinate system. The average incident angle is the arithmetic mean of the angles between the laser incident direction and the normal vector of the point at all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship.

7. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 5, characterized in that, The echo characteristics are used to reflect the quality of the echo signal of the lidar within the variable voxel, including the average echo intensity, the number of echoes, and the echo interval. The average echo intensity is the arithmetic mean of the echo intensity of all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship. The number of echoes is the maximum value of the echo numbers of all discrete points of the lidar under the current changing voxel within the point cloud mapping relationship; The echo interval is the arithmetic mean of the time differences between adjacent echo points of the lidar under the current changing voxel within the point cloud mapping relationship.

8. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 1, characterized in that, The calculation steps for the comprehensive fusion weight of each lidar within the corresponding variable voxel include: Obtain the prior weights of the quality features corresponding to the functional partitions; Based on the quality feature scores of all lidars within the current changing voxel, the data dispersion and information contribution of each quality feature are quantified, and the objective entropy weight of each quality feature is calculated. The prior weights and objective entropy weights are weighted and fused according to a preset weight ratio to obtain the combined weights of each quality feature. For each lidar, its scores for each quality feature within the current changing voxel are multiplied sequentially by the combined weights of the corresponding quality features and then summed to obtain the comprehensive score of the lidar. The overall scores of each lidar are globally normalized to obtain the overall fusion weight.

9. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 1, characterized in that, The steps for generating the fused point cloud map of the current frame include: For each variable voxel, extract the static point cloud data of each participating fusion lidar within that variable voxel from the point cloud mapping relationship, including the three-dimensional coordinates and echo intensity information of each discrete point; Based on the comprehensive fusion weight, the three-dimensional coordinates and echo intensity are weighted and fused separately to obtain the fused coordinates and fused intensity of the changed voxel; The fusion coordinates and fusion intensity of all the variable voxels are stitched together with the fusion coordinates and fusion intensity of the static voxels from the previous frame to obtain the fused point cloud map of the current frame.

10. The multi-radar point cloud weighted fusion method based on entropy weighting as described in claim 1, characterized in that, The step of generating the current frame point cloud data set includes: Responding to the data collection and triggering signals of the intelligent automated warehouse; Based on a unified time reference, the current frame point cloud data of all lidars are collected synchronously. Each lidar-collected point cloud data is preprocessed independently. The preprocessing operations include removing outlier data, filtering out noise data, and performing motion compensation on the point cloud data. All preprocessed point cloud data from LiDAR are uniformly converted to the global coordinate system of the intelligent automated warehouse and integrated to generate the current frame point cloud data set.

11. A multi-radar point cloud weighted fusion system based on entropy weighting method, used to implement the multi-radar point cloud weighted fusion method based on entropy weighting method as described in any one of claims 1-10, characterized in that, include: Data acquisition module: In response to the acquisition trigger signal of the intelligent automated warehouse, it synchronously acquires the point cloud data of all LiDARs in the current frame, performs point cloud data preprocessing, and generates the point cloud data set of the current frame; The static / dynamic separation module is used to perform independent temporal dynamic / static separation of the current frame point cloud data set by a single LiDAR, distinguishing between dynamic and static points, and generating a static point cloud data set. Change detection module: Based on the functional partitioning of fused voxel mesh, it performs single-LiDAR independent adjacent frame change detection on the current frame static point cloud data set, distinguishes between changing voxels and static voxels, and generates a set of voxels to be processed; Feature extraction module: For each variable voxel in the set of voxels to be processed, extract the multi-dimensional quality features of each LiDAR within the variable voxel, and perform differential standardization processing to obtain the quality feature scores of each LiDAR within the variable voxel. Weight calculation module: Based on the functional partition to which the variable voxel belongs, a combined weight is constructed by combining prior weight and entropy weight method. The comprehensive score of each lidar is calculated based on the quality feature score. After normalization, the comprehensive fusion weight of each lidar in the corresponding variable voxel is obtained. Point cloud map generation module: Based on comprehensive fusion weights, it performs incremental voxel-level weighted fusion operations on changing voxels, and combines the fusion results of the previous frame with the static voxels to generate the fused point cloud map of the current frame.

Images